Why can’t neural networks (NN) forecast better? In the major super-forecasting competitions, NN have typically under-performed when compared to traditional statistical methods. When they have performed well, the underlying methods have been ensembles of NN and statistical methods. Forecasting stock markets, medical, infrastructure dynamics, social activity or pandemics each have their own challenges. In this study, we evaluate the strengths of a collection of methods for forecasting pandemics such as Covid-19 using NN, statistical methods as well as parameterized mechanistic models. Forecasts of epidemics can inform public health response and decision making, so accurate forecasting is crucial for general public notification, timing and spatial targeting of intervention. We show that NN typically under-perform in forecasting Covid-19 active cases which can be attributed to the lack of training data which is common for forecasts. Our test data consists of the top ten countries for active Covid-19 cases early in the pandemic and is represented as a Time Series (TS). We found that Statistical methods outperform NN for most cases. Albeit, NN are still good pattern finders and we suggest that there are perhaps more productive ways other than purely
data driven models of using NN to help produce better forecasts.
PDF